LLM - Day 1 - Intro

Large Language Model - Its a model - what is a model ? its a equation. - Then what is equation ? y = mx + c or y = m1x2 + m2x3 + m3x2 - Models are nothing but WEIGHTS. - Take an example Eg: gpt-oss-120b , for this model they would have used data with TRANSFORMER ARCHITECTURE and we are getting some output that is called MODEL. m --> weights ( we can adjust the m and get many desierd outputs ) - Model is floating point storage. - சில பேரு சொல்லுவாங்க இது வந்து 120B Paramter model , what it means ? These Parameters are nothing but how many weights I am having inside the model. நான் எந்த அளவுக்கு dataவை கொடுத்து train பண்ணி இருக்கேன்னு அர்த்தம். In other words , CONTEXT அதிகம் . - Can we run this 120B Paramter in machine ? --> Where the OS sits ? C drive or /boot , initially it will be in hard disc , once the machine starts , it will go to RAM Running memory போய் உட்கார்ந்து கொல்லும். So to run the model we need RAM and also the CPU , so these two decides which model to run. Eg., in laptop we can run 7B parameter. - In local like laptop we can run eg., llama cpp, ollama , LM Studio . இது எல்லாம் modelஆ ? No , these are softwares/tools to run the model. - Open model are called like " Open weights model ". Eg., chatgpt model , gemini model. ( Referring to Open source ) Why I Need a model ? - LLM is a next word predictor. Eg., Hi How ......... இதுக்கப்புறம் என்ன வரலாம் like are you , is your life , is the food . இது எல்லாம் ஒரு possibilities ( It can have n no of possibilites ). MORE THE PRESENCE , MORE POSSIBILITY OF OCCURING. நீங்க ஒரு வார்த்தையை குடுக்குறீங்க , அதுக்கு அடுத்த வார்த்தை இது கண்டுபிடித்து தர போது . Then this will go as input , like wise it will generate. - All each possibilities will have SCORE , entire score will like between [ 0 - 100 ] or [ 0 - 1 ] . - If you are not giving any controlling parameter , then it will chose the highest SCORE. - If you want to control the LLM , at that place . We have controlling parameters. Namely : - Temparature - Top-k - Top-p Temparature - 0-1 - controls how random or safe the model choices. - If the TEMP is low like ( 0-0.3 ) then it will ACT BASED ON THE FACTS. Likely having max score / most occuring one. - If the TEMP is high like ( 0.7 - 1+ ) , model is willing to choice less likely words. Eg., in storing telling or in creating writing. It will choose less likely words from dataset which is trained on. Top-k - Along with temperature its used. - Model only considers , Top k most items. - K --> Elements Top-P - Also its called Sampling methods. - It will add the score and then it will work. - Along with temperature its used. இப்ப Top - pகு 90 % value சொல்லிட்டாங்க . இப்ப அது எப்படி work ஆகும் நா. Each prediction will have a score eg., first prediction 0.70 , அடுத்த prediction 0.15 , அடுத்த prediction 0.30 . It will add the prediction score which is equal to 90 % ( more or less ) இத வச்சு தான் அந்த output கொடுக்கும். Tokens - words ? - which are " Not complete ". - Entire text into pieces. - Tokenization ? Process of chopping your text into tokens. - Each token will be assigned numbers. - Take it like a translator. Context Window - Model ஓட short memory. - Short term memory. - Max of tokens a model can see in a given time. - ஒரு timeல ஒரு model லால , எவ்வளவு விஷயத்தை பார்க்க முடியும். TBD : - What is dataset ? - Chunkings ? Notes : - If I am going to use a model , then I need to unzip it , for this i need a running memory so we need RAM. Now , i need process the data , then CPU & GPU ( training a& usage ) are required. - Weigts will be like FOLATING POINTS or EMBEDDINGS. - https://groq.com/ --> Its a tool --> They give free models for usage. - Fine Tuning is not updating the model or touching the weights. We use like rapper around the model thats called Fine Tunning. - Markov Chain --> this is used in next word prediction in gmail , word etc . - Harry Porter and context window . Top comments (0)